Each node is an 8U system built for large-scale inference serving.
GPU
8 x NVIDIA B300 on the NVIDIA HGX platform. 288 GB each, 2.3 TB per node.
GPU interconnect
NVIDIA NVLink with NVSwitch.
CPU and memory
2 x AMD EPYC 9575F, 128 cores per node, with 3 TB of DDR5-6400.
Local storage
8 x 7.6 TB E1.S NVMe SSD, plus 2 x 1.9 TB M.2 NVMe.
Network
8 x 800GbE (XDR800) and 2 x dual-port 200GbE (NDR200).
Tenancy and security
Single-tenant bare metal. Encryption in transit and at rest. SOC 2 Type II.
The capacity, the environment around it and the people who run it.
Full-node reservation
Your nodes are yours all day, every day, for the term. There is no daily window and no preemption.
Single-tenant bare metal
No other customer's workload runs on your assigned nodes.
Cluster networking
Eight 800 Gb/s ports on every node, one for each GPU, with RDMA across the cluster.
A persistent environment
Storage volumes, images and network configuration stay in place across the term.
Hot-spare node pool
Spare nodes on hand, with a 4-hour replacement target.
A named team
A named technical account manager, a shared Slack channel and 24/7 engineering support.
$8.50 per GPU-hour on NVIDIA B300, on a 12-month minimum agreement. Capacity is reserved for you and billed monthly. Larger commitments and longer terms are quoted by our team.
$8.50
per GPU-hour, 12-month minimum
The same GPUs sit behind per-token inference. Start there and move to dedicated nodes when the workload calls for it.

By the token. Open-weight models behind an OpenAI-compatible API, with no commitment.
Learn more
By the hour. A whole GPU node, self-serve, with no term.
Learn more
Keep datasets, checkpoints and results next to your compute.
Learn more
The open-weight models available on Telnyx.
Learn moreTell us the workload, the GPU count and where it needs to sit. We will come back with a written rate.

Tell us the workload, the GPU count and where it needs to sit. We will come back with a written rate.

Dedicated Compute is reserved, single-tenant NVIDIA B300 capacity on hardware Telnyx owns and operates. You rent the nodes for a term and they are yours around the clock.
$8.50 per GPU-hour on a 12-month minimum agreement. The rate is fixed for the full term.
12 months. Capacity is sold in whole nodes of 8 GPUs. [SMALLEST DEAL SIZE FROM DEAL DESK]
Bare-metal nodes with 8 NVIDIA B300 GPUs at 288 GB each, two AMD EPYC 9575F processors, 3 TB of system memory, 8 x 7.6 TB of local NVMe and eight 800GbE ports.
No. Assigned nodes are single-tenant, with no oversubscription and no preemption.
On Telnyx sites across the Americas, Europe, APAC and MENA. Placement is agreed with you, based on where your data and users are.
We hold a hot-spare node pool and replace a failed node from it, with a 4-hour replacement target.
[ANSWER FROM PRODUCT: bare metal access, Kubernetes or Slurm]